Pith. sign in

REVIEW 4 major objections 4 minor 70 references

Hi-LSplat: Hierarchical 3D Language Gaussian Splatting

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Hi-LSplat builds a view-consistent 3D semantic tree that answers open-vocabulary queries from whole objects down to their parts.

desk verdict Plausible hierarchical language Gaussian work with strong benchmark results, but the central view-consistency claim is not quantitatively supported and several tables have mechanical problems. read the letter →

arxiv 2506.06822 v1 pith:WG6NUNVY submitted 2025-06-07 cs.CV cs.AI

Chenlu Zhan , Yufei Zhang , Gaoang Wang , Hongwei Wang This is my paper
classification cs.CVcs.AI
keywords hierarchicallanguageGaussiansplattingopen-vocabulary3Dsegmentationfieldsview-consistentsemanticscontrastivelearningSAMsemanticmaskstreelocalization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Hi-LSplat sets out to show that open-vocabulary 3D scene understanding works better when the semantic hierarchy—whole, part, subpart—is built once in 3D instead of assembled from per-view 2D segmentations. The method constructs a 3D hierarchical semantic tree from SAM masks at three granularities, then trains language features so that each tree node is view-consistent and clearly separated from its siblings. If the central claim holds, a single 3D language field can answer queries like "bear nose" or "reflective part of stainless steel" with masks that agree across viewpoints. The paper reports improvements of 34.14 and 8.0 mIoU on ScanNet and LERF, and contributes two new hierarchical datasets for evaluating this ability.

What carries the argument

The load-bearing object is the 3D hierarchical semantic tree, built by first using SAM with a 32 by 32 point-prompt grid to produce whole, part, and subpart masks per view; a coverage threshold theta = 0.9 turns overlaps between these masks into parent-child relations, yielding three semantic levels. Rendered 3D Gaussian features inside each boolean mask are clustered to the mask's mean feature, which forms a tree node, and the hierarchical clustering loss enforces that consistency in 3D space. Two contrastive losses operate on that tree: the instance-wise loss pushes different tree nodes apart with margins scaled by the semantic-level distance, and the part-wise loss computes similarities after removing the shared parent feature, so sibling subparts are separated without being compared through their common semantics.

What would settle it

Render each node of the trained 3D tree back into the training views and measure overlap between the projected node region and that view's original SAM mask; if a part node such as "bear nose" misses its per-view mask in more than about ten percent of views, the view-consistency claim fails. A second check is to compare tree nesting against manually annotated part relationships on the new Hi-LERF and Hi-3DOVS datasets to see whether the theta = 0.9 parent-child assignments match the ground-truth whole-part-subpart structure.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the root cause of inconsistency in previous language-Gaussian models is that they inherit view-dependent 2D features instead of learning a unified 3D representation. Hi-LSplat therefore replaces 2D pixel-aligned features with a 3D hierarchical semantic tree: per-view SAM boolean masks at three levels are converted into per-mask mean features, an overlap threshold (theta = 0.9) decides which mask contains which, and a point-optimized hierarchical clustering loss pulls rendered 3D features toward those means. Instance-wise and part-wise contrastive losses then shape the tree so that different objects are pushed apart by margins that depend on their semantic level, while sibling parts are separated after subtracting their shared parent feature. The result is a field that the paper argues captures all-sided hierarchical semantics and outperforms prior methods on 3D open-vocabulary segmentation and localization across eight datasets.

Load-bearing premise

The argument rests on the assumption that SAM masks generated at three prompt scales, linked by the theta = 0.9 overlap rule, define the same correct whole-part-subpart hierarchy in every view and that clustering 3D features to per-mask means preserves that hierarchy.

Editorial extensions

If this is right

  • A user can query the same 3D scene at different granularities—"bear", "bear nose", "bear nostril"—and each query activates the corresponding tree level rather than a single blended feature.
  • Part-level masks become stable across viewpoints because the hierarchy is anchored to 3D cluster means, not to per-view 2D segmentations.
  • The instance-wise and part-wise contrastive losses separate semantically close siblings (nose vs. mouth) and objects at different levels (stuffed bear vs. bear nose) in one training run.
  • The two new hierarchical datasets Hi-LERF and Hi-3DOVS supply per-level IoU and a hierarchical consistency score, so future methods can be compared on part-level semantics instead of only whole-object queries.
  • If the method generalizes as claimed, downstream consumers such as robotic pick-and-place or 3D editing can select parts directly from the tree rather than re-segmenting per frame.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next step the paper leaves implicit is to answer free-form relational queries ("the cup on the white sink") by traversing the tree instead of matching text to a single node.
  • The hierarchical consistency score could plausibly be used as a training regularizer, which might reduce sensitivity to the overlap threshold theta and to SAM mask noise.
  • One testable extension is to make theta adaptive per object or per view, since thin or reflective structures are where SAM masks are most likely to disagree across viewpoints.
  • The same tree structure may support part-aware 3D editing, such as recoloring or moving one part while preserving its parent, a consequence the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Hi-LSplat, a 3D Gaussian Splatting language field that augments each Gaussian with a learnable semantic feature and organizes the scene into a hierarchical semantic tree with three levels (whole, part, subpart). The hierarchy is derived from SAM masks per training view, with overlap-based parent-child assignment using a coverage threshold theta, and is optimized with a hierarchical clustering loss, an instance-wise contrastive loss, and a part-wise contrastive loss. The authors also construct two hierarchical datasets, Hi-LERF and Hi-3DOVS, by annotating LERF and 3D-OVS frames with three levels of masks. The paper reports open-vocabulary segmentation, localization, and hierarchy-consistency results on 8 datasets and claims state-of-the-art performance, including gains of 34.14 and 8.0 mIoU on ScanNet and LERF.

Significance. If the claims were fully substantiated, the paper would make a useful contribution to open-vocabulary 3D scene understanding: hierarchical part-level queries such as 'bear nose' are an important step beyond object-level language fields, and the two new datasets could benefit the community. The method is plausible and builds naturally on existing components (SAM, Gaussian Splatting, CLIP). However, the quantitative evidence has several load-bearing inconsistencies: LERF baseline averages are computed with missing scenes as zero, the reported average for the proposed method does not match the per-scene numbers, and the headline improvement values are not traceable to any table. The view-consistency claim, which is the paper's central conceptual contribution, is supported only qualitatively.

major comments (4)
  1. [III-B, Eq. (2), and Fig. 11] The view-consistency claim is not established. Section III-B says 'for each viewpoint, we generate the corresponding 2D masks' and at the same time calls these masks 'view-independent'; these statements are in direct tension. The hierarchical clustering loss in Eq. (2) and the contrastive losses in Eqs. (3) and (5) are defined on per-view mask means and mask sets, without any explicit cross-view correspondence or a global 3D tree being assembled from the per-view hierarchies. The only consistency evidence is the qualitative 'bear nose' comparison in Fig. 11; no quantitative cross-view metric is reported. The HC score in Eq. (7) measures nesting of annotation masks within a view, not agreement of the same semantic across views. Please add a quantitative view-consistency evaluation, for example mask IoU or feature agreement for the same 3D region across held-out views, and clarify the mechanism by which conflicting per-view mask hierarchies are reconciled into one 3D tree.
  2. [Table III] The LERF averages for FreeGS and Gaussian Grouping are computed by treating the missing Waldo kitchen score as 0: (62.6+77.5+68.5+0)/4=52.2 for FreeGS and (69.7+77.0+71.7+0)/4=54.6 for Gaussian Grouping. This artificially lowers the baselines. If the missing scene is excluded from the average, FreeGS becomes 69.5 and Gaussian Grouping becomes 72.8, both of which are above the reported Ours average of 68.4. In addition, the Ours mIoU average of 68.4 does not match the arithmetic mean of the four listed per-scene values, which is 70.7. The handling of missing scenes must be clarified and all averages recomputed; the LERF SOTA claim currently depends on this inconsistent computation.
  3. [Abstract and Contributions] The claimed improvements of '34.14 and 8.0 mIoU on ScanNet and LERF datasets' are not traceable to any result in the paper. In Table VI, the largest ScanNet 19-class gap over reported baselines is 17.8 mIoU over OpenGaussian (47.9 vs. 30.1), and no LERF comparison in Table III shows a gap of 8.0 mIoU. The abstract and the contributions bullet should be revised to state the actual differences, or the unsubstantiated numbers should be removed.
  4. [Section IV and Table VIII] The two new datasets are annotated with the same SAM-mask and overlap-threshold pipeline that the method itself uses. Section IV states: 'We used SAM to extract hierarchical semantic masks, analyzed their overlaps for semantic layering, and manually annotated hierarchical semantic features with labels as ground truth.' The 3D hierarchical tree in the method is trained to imitate exactly this 2D hierarchy, so the high HC scores on Hi-LERF and Hi-3DOVS partly measure consistency with the construction procedure rather than with an independent semantic hierarchy. Please either use an independent annotation protocol, or report quantitative results on an external hierarchical benchmark, and discuss this limitation explicitly.
minor comments (4)
  1. [Table II] The header of Table II reads 'Hi-LERF' but the contents list Hi-3DOVS scenes such as 'bed', 'bench', and 'bluesofa'; the header should be corrected to 'Hi-3DOVS'.
  2. [Eq. (4)] In the definition of s_p, the denominator contains the same term twice, || M_j^l - M^{l-1} || multiplied by itself; this is likely a typo and should presumably be the product of the norms of the two residual vectors.
  3. [Table IV and Table VIII] There are citation inconsistencies: Table IV lists 'Gaussian Grouping [49]' but Gaussian Grouping is reference [37], and Table VIII lists 'VCH [47]' while the text and reference list identify VCH as [52].
  4. [III-B] The notation M^l is introduced after Eq. (1) defines a single rendered map M; the paper should explicitly state that three hierarchical maps M^w, M^p, and M^s are rendered and how they correspond to the levels l=1,2,3.

Circularity Check

2 steps flagged · score 4.0 of 10

Hierarchical datasets and hyperparameters are self-referential; public-benchmark gains remain externally valid.

  1. self definitional [Section IV (Hierarchical Datasets) vs. Section III-B (3D Hierarchical Semantic Tree)]
    "We used both automatic and manual labeling methods. We used SAM to extract hierarchical semantic masks, analyzed their overlaps for semantic layering, and manually annotated hierarchical semantic features with labels as ground truth. ... Following the 3D semantic hierarchy tree, smaller masks represent higher semantic levels, and they are precisely nested within the previous layer’s tree node masks, ensuring M_ls ⊂ M_lp ⊂ M_lw."

    The ground-truth hierarchy for Hi-LERF and Hi-3DOVS is produced by the same SAM mask extraction plus overlap-threshold analysis that Section III-B uses to construct the method's own 3D hierarchical tree and to define its clustering and contrastive targets (Eqs. 2-5). The dataset description explicitly says it 'follows the 3D semantic hierarchy tree' and enforces exactly the nesting M_s ⊂ M_p ⊂ M_w that the method is trained to reproduce. Therefore the reported HC and mIoU on these two datasets largely measure self-consistency with the paper's own SAM-overlap pipeline rather than agreement with an independently defined semantic hierarchy.

  2. fitted input called prediction [Section V-D (Parameter Discussions, Tables XII and XIV); Section V-A3 (Implementation Details)]
    "The θ and Ω are set 0.9 and 10. ... The best results are obtained when the coverage threshold is set to 0.9. ... We observe that our method achieves optimal performance when Ω=10."

    Tables XII and XIV sweep θ on 3D-OVS (Overall) and Hi-LERF, and Ω on LERF (Average) and Hi-LERF, which are the same datasets whose numbers appear in the main comparison tables. The final settings θ=0.9 and Ω=10 are the values that maximize those very reported metrics. Thus the headline mIoU, HC, and localization numbers on these benchmarks are partly a selected maximum over hyperparameters tuned on the test sets, making the claimed predictive superiority on those datasets statistically forced rather than an independent evaluation.

full rationale

The paper's core comparison on public benchmarks (ScanNet, LERF, 3D-OVS, Replica, Waymo, ScanNet200) is externally grounded and not circular; the clustering and contrastive losses are trained against SAM-derived targets in a straightforward supervised manner. However, the two newly constructed hierarchical datasets and the Hierarchical Consistency metric inherit the same SAM + overlap-threshold construction (θ=0.9) that defines the method's own semantic tree, so the hierarchical-semantics evaluation is partially self-referential. In addition, the two key hyperparameters θ and Ω are selected by sweeping on the same datasets used for the reported results, which statistically forces part of the measured improvement. These issues affect the strength of the hierarchical and consistency claims, but they do not invalidate the independent public-benchmark evidence, so the overall circularity is partial rather than total.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central machinery is built on standard 3DGS alpha-blending plus several domain assumptions about SAM masks, CLIP embeddings, and the new HC metric. The main free parameters are theta, Omega, lambda1, lambda2, the inference threshold, and the unstated temperature tau. No invented physical entities are introduced; the hierarchical tree is a data structure, not a new mediator or force.

free parameters (5)
  • coverage threshold theta = 0.9
    Sets the part-whole inclusion criterion for SAM masks (Section III-B); the ablation in Table XII selects 0.9 on the same datasets used for final reporting.
  • similarity degree Omega = 10
    Target distance ratio between adjacent semantic levels in Eq. (3); chosen via ablation on LERF and Hi-LERF in Table XIV.
  • loss weights lambda1 and lambda2 = 1e-6 and 1e-5
    Weights for instance-wise and part-wise losses in Eq. (6); no ablation or validation protocol is given for these values.
  • inference relevancy threshold = 0.4
    Binarizes relevancy maps for LERF/Hi-LERF segmentation (Section V-A.3); no separate validation split is described.
  • part-wise temperature tau = not reported
    Temperature in Eq. (5) is never specified in the implementation details, so the part-wise contrastive loss is not exactly reproducible.
assumptions (5)
  • standard math Alpha-blending of per-Gaussian features (Eq. 1) is a differentiable and valid image-space summary for 3D language fields.
    Adopted from 3D Gaussian Splatting [13] and prior language-Gaussian methods [4], [9].
  • domain assumption SAM boolean masks extracted per view at three prompt scales define a semantically meaningful whole-part-subpart hierarchy under the theta-overlap rule.
    Used as the initial 2D hierarchy in Section III-B; if masks are not semantically stable or view-consistent, the 3D tree inherits their noise.
  • domain assumption Mean-pooled rendered features inside each per-view boolean mask are compatible optimization targets that converge to one view-consistent 3D instance feature.
    The clustering loss in Eq. (2) assumes per-view mask averages can be unified without explicit multi-view mask association.
  • domain assumption CLIP image and text embeddings preserve hierarchical and part-level semantic distinctions for open-vocabulary queries.
    Inference in Section III-D relies on CLIP alignment and a per-scene decoder to map 3-D latent features back to CLIP space.
  • ad hoc to paper The hierarchical consistency score in Eq. (7) is a valid measure of part-whole nesting quality.
    Introduced by the authors and not validated against external benchmarks or human judgments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hi-LSplat: Hierarchical 3D Language Gaussian Splatting." pith.science (2026). https://pith.science/paper/WG6NUNVY

@misc{pith2026250606822,
  author       = {Pith},
  title        = {Pith review of: Hi-LSplat: Hierarchical 3D Language Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WG6NUNVY}},
  note         = {Machine review of arXiv:2506.06822}
}
read the original abstract

Modeling 3D language fields with Gaussian Splatting for open-ended language queries has recently garnered increasing attention. However, recent 3DGS-based models leverage view-dependent 2D foundation models to refine 3D semantics but lack a unified 3D representation, leading to view inconsistencies. Additionally, inherent open-vocabulary challenges cause inconsistencies in object and relational descriptions, impeding hierarchical semantic understanding. In this paper, we propose Hi-LSplat, a view-consistent Hierarchical Language Gaussian Splatting work for 3D open-vocabulary querying. To achieve view-consistent 3D hierarchical semantics, we first lift 2D features to 3D features by constructing a 3D hierarchical semantic tree with layered instance clustering, which addresses the view inconsistency issue caused by 2D semantic features. Besides, we introduce instance-wise and part-wise contrastive losses to capture all-sided hierarchical semantic representations. Notably, we construct two hierarchical semantic datasets to better assess the model's ability to distinguish different semantic levels. Extensive experiments highlight our method's superiority in 3D open-vocabulary segmentation and localization. Its strong performance on hierarchical semantic datasets underscores its ability to capture complex hierarchical semantics within 3D scenes.

Figures

Figures reproduced from arXiv: 2506.06822 by the authors.

Figure 2
Figure 2. Comparison of 3D open-vocabulary semantic query and semantic feature between our model and SOTA hierarchical semantic models. We highlighted [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The network structure of Hi-LSplat. We propose a view-consistent hierarchical language Gaussian Splatting work for 3D open-vocabulary querying [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. We randomly selected several open-vocabulary queries and their cor [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Comparison of open-vocabulary semantic query (left) and semantic localization task (right) on the LERF dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: More comparison of open-vocabulary semantic query on the LERF dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: More comparison of open-vocabulary semantic query on the 3D-OVS and Hi-3DOVS datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Comparison of open-vocabulary semantic query on the “sofa” scene of 3D-OVS and Hi-3DOVS datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Comparison of hierarchy on the Hi-LERF dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Comparison of hierarchy semantic features. Views:13 Views: 52 Views:1 [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Comparison of consist semantics of “bear nose”. [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 40 canonical work pages

  1. [1]

    Simvqa: Exploring simulated environments for visual question answer- ing,

    P. Cascante-Bonilla, H. Wu, L. Wang, R. S. Feris, and V . Ordonez, “Simvqa: Exploring simulated environments for visual question answer- ing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 5056–5066

    work page 2022

  2. [2]

    Scanqa: 3d question answering for spatial scene understanding,

    D. Azuma, T. Miyanishi, S. Kurita, and M. Kawanabe, “Scanqa: 3d question answering for spatial scene understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 19 129–19 139

    work page 2022

  3. [3]

    Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning,

    H. Ying, Y . Yin, J. Zhang, F. Wang, T. Yu, R. Huang, and L. Fang, “Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 612–20 622

    work page 2024

  4. [4]

    Langsplat: 3d language gaussian splatting,

    M. Qin, W. Li, J. Zhou, H. Wang, and H. Pfister, “Langsplat: 3d language gaussian splatting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 051–20 060

    2024

  5. [5]

    Sai3d: Segment any instance in 3d scenes,

    Y . Yin, Y . Liu, Y . Xiao, D. Cohen-Or, J. Huang, and B. Chen, “Sai3d: Segment any instance in 3d scenes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 3292–3302

    work page 2024

  6. [6]

    Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip,

    Q. Yu, J. He, X. Deng, X. Shen, and L.-C. Chen, “Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip,” inAdvances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 32 215–32 234. [Online]. Available: https://...

    work page 2023

  7. [7]

    Weakly supervised 3d open-vocabulary segmentation,

    K. Liu, F. Zhan, J. Zhang, M. XU, Y . Yu, A. El Saddik, C. Theobalt, E. Xing, and S. Lu, “Weakly supervised 3d open-vocabulary segmentation,” inAdvances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 53 433–53 456. [Online]. Available: https:/...

    work page 2023

  8. [8]

    Visual language maps for robot navigation,

    C. Huang, O. Mees, A. Zeng, and W. Burgard, “Visual language maps for robot navigation,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 10 608–10 615

    2023

Show all 70 references
  1. [9]

    Opengaussian: Towards point- level 3d gaussian-based open vocabulary understanding,

    Y . Wu, J. Meng, H. Li, C. Wu, Y . Shi, X. Cheng, C. Zhao, H. Feng, E. Ding, J. Wanget al., “Opengaussian: Towards point- level 3d gaussian-based open vocabulary understanding,”arXiv preprint arXiv:2406.02058, 2024

    2024 arXiv

  2. [10]

    Gsnerf: Gen- eralizable semantic neural radiance fields with enhanced 3d scene un- derstanding,

    Z.-T. Chou, S.-Y . Huang, I. Liu, Y .-C. F. Wanget al., “Gsnerf: Gen- eralizable semantic neural radiance fields with enhanced 3d scene un- derstanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 806–20 815

    2024

  3. [11]

    Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understand- ing,

    X. Zuo, P. Samangouei, Y . Zhou, Y . Di, and M. Li, “Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understand- ing,”arXiv preprint arXiv:2401.01970, 2024

    2024 arXiv

  4. [12]

    Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,

    D. Chen, H. Li, W. Ye, Y . Wang, W. Xie, S. Zhai, N. Wang, H. Liu, H. Bao, and G. Zhang, “Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,”IEEE Transactions on Visualization and Computer Graphics, 2024

    2024

  5. [13]

    3d gaussian splatting for real-time radiance field rendering,

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering,”ACM Transactions on Graphics, vol. 42, no. 4, 2023

    2023

  6. [14]

    Aligning instance-semantic sparse representation towards unsupervised object segmentation and shape abstraction with repeatable primitives,

    J. Li, H. Wang, J. Tan, Z. Ou, and J. Yuan, “Aligning instance-semantic sparse representation towards unsupervised object segmentation and shape abstraction with repeatable primitives,”IEEE Transactions on Visualization and Computer Graphics, 2025

    2025

  7. [15]

    Jimr: Joint semantic and geometry learning for point scene instance mesh reconstruction,

    Q. Yu, X. Li, Y . Tang, J. Xu, L. Hu, Y . Hao, and M. Chen, “Jimr: Joint semantic and geometry learning for point scene instance mesh reconstruction,”IEEE Transactions on Visualization and Computer Graphics, 2024

    2024

  8. [16]

    Delving into invisible semantics for generalized one-shot neural human rendering,

    Y . Lin, X. Xu, H. Zhang, C. Xu, W. Li, Y . Xie, J. Qin, and S. He, “Delving into invisible semantics for generalized one-shot neural human rendering,”IEEE Transactions on Visualization and Computer Graphics, 2025

    2025

  9. [17]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” 2021. [Online]. Available: https://arxiv.org/abs/2103.00020

    2021 arXiv

  10. [18]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Dollar, and R. Girshick, “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 4015–4026

    2023

  11. [19]

    Sam3d: Segment anything in 3d scenes,

    Y . Yang, X. Wu, T. He, H. Zhao, and X. Liu, “Sam3d: Segment anything in 3d scenes,” 2023. [Online]. Available: https: //arxiv.org/abs/2306.03908

    2023 arXiv

  12. [20]

    Segment any 3d gaussians,

    J. Cen, J. Fang, C. Yang, L. Xie, X. Zhang, W. Shen, and Q. Tian, “Segment any 3d gaussians,” 2024. [Online]. Available: https://arxiv.org/abs/2312.00860

    2024 arXiv

  13. [21]

    Click-gaussian: Interactive segmentation to any 3d gaussians,

    S. Choi, H. Song, J. Kim, T. Kim, and H. Do, “Click-gaussian: Interactive segmentation to any 3d gaussians,” 2024. [Online]. Available: https://arxiv.org/abs/2407.11793

    2024 arXiv

  14. [22]

    Dino: Detr with improved denoising anchor boxes for end-to-end object detection,

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. M. Ni, and H.-Y . Shum, “Dino: Detr with improved denoising anchor boxes for end-to-end object detection,” 2022. [Online]. Available: https://arxiv.org/abs/2203.03605

    2022 arXiv

  15. [23]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  16. [24]

    Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding,

    L. Xue, M. Gao, C. Xing, R. Mart ´ın-Mart´ın, J. Wu, C. Xiong, R. Xu, J. C. Niebles, and S. Savarese, “Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding,” 2023. [Online]. Available: https://arxiv.org/abs/2212.05171

    2023 arXiv

  17. [25]

    Clip-gs: Clip-informed gaussian splatting for real-time and view- consistent 3d semantic understanding,

    G. Liao, J. Li, Z. Bao, X. Ye, J. Wang, Q. Li, and K. Liu, “Clip-gs: Clip-informed gaussian splatting for real-time and view- consistent 3d semantic understanding,” 2024. [Online]. Available: https://arxiv.org/abs/2404.14249

    2024 arXiv

  18. [26]

    Fastlgs: Speeding up language embedded gaussians with feature grid mapping,

    Y . Ji, H. Zhu, J. Tang, W. Liu, Z. Zhang, Y . Xie, and X. Tan, “Fastlgs: Speeding up language embedded gaussians with feature grid mapping,”

  19. [27]

    Pla: Language- driven open-vocabulary 3d scene understanding,

    R. Ding, J. Yang, C. Xue, W. Zhang, S. Bai, and X. Qi, “Pla: Language- driven open-vocabulary 3d scene understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 7010–7019

    2023

  20. [28]

    Openmask3d: Open-vocabulary 3d instance segmentation,

    A. Takmaz, E. Fedele, R. W. Sumner, M. Pollefeys, F. Tombari, and F. Engelmann, “Openmask3d: Open-vocabulary 3d instance segmentation,” 2023. [Online]. Available: https://arxiv.org/abs/2306. 13631

    2023

  21. [29]

    Jiang, A. Tagliasacchi, M. Pollefeys, and T. Funkhouser, “Openscene: 3d scene understanding with open vocabu- laries,

    S. Peng, K. Genova, C. “. Jiang, A. Tagliasacchi, M. Pollefeys, and T. Funkhouser, “Openscene: 3d scene understanding with open vocabu- laries,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 815–824

    2023

  22. [30]

    Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting,

    J. Guo, X. Ma, Y . Fan, H. Liu, and Q. Li, “Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting,”

  23. [31]

    Gaussianeditor: Editing 3d gaussians delicately with text instructions,

    J. Wang, J. Fang, X. Zhang, L. Xie, and Q. Tian, “Gaussianeditor: Editing 3d gaussians delicately with text instructions,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 20 902–20 911

    2024

  24. [33]

    Language embedded 3d gaussians for open-vocabulary scene understanding,

    J.-C. Shi, M. Wang, H.-B. Duan, and S.-H. Guan, “Language embedded 3d gaussians for open-vocabulary scene understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 5333–5343

    2024

  25. [34]

    Editanything: Empowering unparalleled flexibility in image editing and generation,

    S. Gao, Z. Lin, X. Xie, P. Zhou, M.-M. Cheng, and S. Yan, “Editanything: Empowering unparalleled flexibility in image editing and generation,” inProceedings of the 31st ACM International Conference on Multimedia, ser. MM ’23. New York, NY , USA: Association for Computing Machi...

    2023

  26. [35]

    Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields,

    S. Zhou, H. Chang, S. Jiang, Z. Fan, Z. Zhu, D. Xu, P. Chari, S. You, Z. Wang, and A. Kadambi, “Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp....

    2024

  27. [36]

    Lerf: Language embedded radiance fields,

    J. Kerr, C. M. Kim, K. Goldberg, A. Kanazawa, and M. Tancik, “Lerf: Language embedded radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 729–19 739

    2023

  28. [37]

    Gaussian grouping: Segment and edit anything in 3d scenes,

    M. Ye, M. Danelljan, F. Yu, and L. Ke, “Gaussian grouping: Segment and edit anything in 3d scenes,” 2024. [Online]. Available: https://arxiv.org/abs/2312.00732

    2024 arXiv

  29. [38]

    Contrastive gaussian clustering: Weakly supervised 3d scene segmentation,

    M. C. Silva, M. Dahaghin, M. Toso, and A. D. Bue, “Contrastive gaussian clustering: Weakly supervised 3d scene segmentation,” 2024. [Online]. Available: https://arxiv.org/abs/2404.12784

    2024 arXiv

  30. [39]

    Open-vocabulary semantic segmentation with mask-adapted clip,

    F. Liang, B. Wu, X. Dai, K. Li, Y . Zhao, H. Zhang, P. Zhang, P. Vajda, and D. Marculescu, “Open-vocabulary semantic segmentation with mask-adapted clip,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 7061–7070

    2023

  31. [40]

    Tuning-free universally-supervised semantic segmentation,

    X. Yang and X. Gong, “Tuning-free universally-supervised semantic segmentation,” 2024. [Online]. Available: https://arxiv.org/abs/2405. 14294

    2024

  32. [41]

    Sohes: Self-supervised open- world hierarchical entity segmentation,

    S. Cao, J. Gu, J. Kuen, H. Tan, R. Zhang, H. Zhao, A. Nenkova, L.-Y . Gui, T. Sun, and Y .-X. Wang, “Sohes: Self-supervised open- world hierarchical entity segmentation,” 2024. [Online]. Available: https://arxiv.org/abs/2404.12386

    2024

  33. [42]

    Spher- ical mask: Coarse-to-fine 3d point cloud instance segmentation with spherical representation,

    S. Shin, K. Zhou, M. Vankadari, A. Markham, and N. Trigoni, “Spher- ical mask: Coarse-to-fine 3d point cloud instance segmentation with spherical representation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 4060–4069

    2024

  34. [43]

    Hierarchical aggre- gation for 3d instance segmentation,

    S. Chen, J. Fang, Q. Zhang, W. Liu, and X. Wang, “Hierarchical aggre- gation for 3d instance segmentation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021, pp. 15 467–15 476

    2021

  35. [44]

    Bsp-net: Generating compact meshes via binary space partitioning,

    Z. Chen, A. Tagliasacchi, and H. Zhang, “Bsp-net: Generating compact meshes via binary space partitioning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  36. [45]

    Hierarchical open-vocabulary universal image segmentation,

    X. Wang, S. Li, K. Kallidromitis, Y . Kato, K. Kozuka, and T. Darrell, “Hierarchical open-vocabulary universal image segmentation,” in Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran ...

    2023

  37. [46]

    Hihpq: Hierarchical hyperbolic product quantization for unsupervised image retrieval,

    Z. Qiu, J. Liu, Y . Chen, and I. King, “Hihpq: Hierarchical hyperbolic product quantization for unsupervised image retrieval,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 5, pp. 4614–4622, Mar. 2024. [Online]. Available: https://ojs.aaai.org/ind...

    2024

  38. [47]

    Hierarchical intra- modal correlation learning for label-free 3d semantic segmentation,

    X. Kang, L. Chu, J. Li, X. Chen, and Y . Lu, “Hierarchical intra- modal correlation learning for label-free 3d semantic segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 28 244–28 253

    2024

  39. [49]

    Garfield: Group anything with radiance fields,

    C. M. Kim, M. Wu, J. Kerr, K. Goldberg, M. Tancik, and A. Kanazawa, “Garfield: Group anything with radiance fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 21 530–21 539

    2024

  40. [50]

    Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding,

    K. Mo, S. Zhu, A. X. Chang, L. Yi, S. Tripathi, L. J. Guibas, and H. Su, “Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  41. [51]

    Maskgroup: Hi- erarchical point grouping and masking for 3d instance segmentation,

    M. Zhong, X. Chen, X. Chen, G. Zeng, and Y . Wang, “Maskgroup: Hi- erarchical point grouping and masking for 3d instance segmentation,” in 2022 IEEE International Conference on Multimedia and Expo (ICME), 2022, pp. 1–6

    2022

  42. [52]

    Hierarchical point- based active learning for semi-supervised point cloud semantic segmen- tation,

    Z. Xu, B. Yuan, S. Zhao, Q. Zhang, and X. Gao, “Hierarchical point- based active learning for semi-supervised point cloud semantic segmen- tation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 18 098–18 108

    2023

  43. [53]

    Unsupervised hyperbolic metric learning,

    J. Yan, L. Luo, C. Deng, and H. Huang, “Unsupervised hyperbolic metric learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 12 465–12 474

    2021

  44. [54]

    View-consistent hierarchical 3d segmentation using ultrametric feature fields,

    H. He, C. Stearns, A. W. Harley, and L. J. Guibas, “View-consistent hierarchical 3d segmentation using ultrametric feature fields,” 2024. [Online]. Available: https://arxiv.org/abs/2405.19678

    2024 arXiv

  45. [55]

    Laser: Efficient language-guided segmentation in neural radiance fields,

    X. Miao, H. Duan, Y . Bai, T. Shah, J. Song, Y . Long, R. Ranjan, and L. Shao, “Laser: Efficient language-guided segmentation in neural radiance fields,”arXiv preprint arXiv:2501.19084, 2025

    2025 arXiv

  46. [56]

    Language-driven semantic segmentation,

    B. Li, K. Q. Weinberger, S. Belongie, V . Koltun, and R. Ranftl, “Language-driven semantic segmentation,” 2022. [Online]. Available: https://arxiv.org/abs/2201.03546

    2022 arXiv

  47. [57]

    Rethinking open-vocabulary segmentation of radiance fields in 3d space,

    H. Lee, Y . Yun, J. Bae, S. Kim, and Y . Uh, “Rethinking open-vocabulary segmentation of radiance fields in 3d space,” 2025. [Online]. Available: https://arxiv.org/abs/2408.07416

    2025 arXiv

  48. [58]

    Slag: Scalable language- augmented gaussian splatting,

    L. Szilagyi, F. Engelmann, and J. Bohg, “Slag: Scalable language- augmented gaussian splatting,”IEEE Robotics and Automation Letters, 2025

    2025

  49. [59]

    Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting,

    J. Guo, X. Ma, Y . Fan, H. Liu, and Q. Li, “Semantic gaussians: Open-vocabulary scene understanding with 3d gaussian splatting,”arXiv preprint arXiv:2403.15624, 2024

    2024 arXiv

  50. [60]

    Supergseg: Open-vocabulary 3d segmentation with structured super-gaussians,

    S. Liang, S. Wang, K. Li, M. Niemeyer, S. Gasperini, N. Navab, and F. Tombari, “Supergseg: Open-vocabulary 3d segmentation with structured super-gaussians,”arXiv preprint arXiv:2412.10231, 2024

    2024

  51. [61]

    Bootstraping clustering of gaussians for view-consistent 3d scene understanding,

    W. Zhang, L. Zhang, P. Hu, L. Ma, Y . Zhuge, and H. Lu, “Bootstraping clustering of gaussians for view-consistent 3d scene understanding,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 10, 2025, pp. 10 166–10 175

    2025

  52. [62]

    Fastlgs: Speeding up language embedded gaussians with feature grid mapping,

    Y . Ji, H. Zhu, J. Tang, W. Liu, Z. Zhang, X. Tan, and Y . Xie, “Fastlgs: Speeding up language embedded gaussians with feature grid mapping,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 3922–3930

    2025

  53. [63]

    Decomposing nerf for editing via feature field distillation,

    S. Kobayashi, E. Matsumoto, and V . Sitzmann, “Decomposing nerf for editing via feature field distillation,”Advances in Neural Information Processing Systems, vol. 35, pp. 23 311–23 330, 2022

    2022

  54. [64]

    Open- vocabulary panoptic segmentation with text-to-image diffusion models,

    J. Xu, S. Liu, A. Vahdat, W. Byeon, X. Wang, and S. De Mello, “Open- vocabulary panoptic segmentation with text-to-image diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 2955–2966

    2023

  55. [65]

    The replica dataset: A digital replica of indoor spaces,

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Vermaet al., “The replica dataset: A digital replica of indoor spaces,”arXiv preprint arXiv:1906.05797, 2019

    1906 arXiv

  56. [66]

    Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understand- ing,

    X. Zuo, P. Samangouei, Y . Zhou, Y . Di, and M. Li, “Fmgs: Foundation model embedded 3d gaussian splatting for holistic 3d scene understand- ing,”International Journal of Computer Vision, vol. 133, no. 2, pp. 611–627, 2025

    2025

  57. [67]

    Language-grounded indoor 3d semantic segmentation in the wild,

    D. Rozenberszki, O. Litany, and A. Dai, “Language-grounded indoor 3d semantic segmentation in the wild,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 125–141

    2022

  58. [68]

    Scannet: Richly-annotated 3d reconstructions of indoor 15 scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor 15 scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5828–5839

    2017

  59. [69]

    Scalability in perception for autonomous driving: Waymo open dataset,

    P. Sun, H. Kretzschmar, X. Dotiwalla, A. Chouard, V . Patnaik, P. Tsui, J. Guo, Y . Zhou, Y . Chai, B. Caineet al., “Scalability in perception for autonomous driving: Waymo open dataset,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020...

    2020

  60. [70]

    Dr. splat: Directly referring 3d gaussian splatting via direct language embedding registration,

    K. Jun-Seong, G. Kim, K. Yu-Ji, Y .-C. F. Wang, J. Choe, and T.-H. Oh, “Dr. splat: Directly referring 3d gaussian splatting via direct language embedding registration,”arXiv preprint arXiv:2502.16652, 2025

    2025 arXiv

  61. [72]

    Laion-400m: Open dataset of clip-filtered 400 million image-text pairs,

    C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki, “Laion-400m: Open dataset of clip-filtered 400 million image-text pairs,” 2021. [Online]. Available: https://arxiv.org/abs/2111.02114

    2021 arXiv

  62. [2024]

    Available: https://arxiv.org/abs/2406.01916 14

    [Online]. Available: https://arxiv.org/abs/2406.01916 14

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.